SAMONAI: Zero-shot 3D MRI Segmentation

• SAMONAI is a zero-shot, 3D prompt propagation algorithm that leverages the 2D Segment Anything Model to achieve consistent volumetric organ segmentation in MRI.
• It implements a three-step pipeline—point-prompt initialization, orthogonal view seeding, and full-volume fusion—to enhance segmentation accuracy, with improvements up to 11% in Dice scores.
• The approach inspires broader SAM-centric frameworks in weak supervision and encoder adaptation, extending its utility to varied imaging and segmentation tasks.

SAMONAI denotes, most explicitly, a zero-shot, 3D prompt propagation algorithm that leverages the Segment Anything Model to segment 3D regions of interest from a single point prompt in volumetric MRI, and was introduced within an automated framework for surgical outcome prediction in colorectal liver metastasis (CRLM) from pre- and post-contrast MRI (Alberb et al., 10 Sep 2025). In adjacent literature, the same label is also used more loosely as a SAM-centric theme for injecting object-aware priors into downstream segmentation pipelines, including weakly supervised semantic segmentation and foundation-model adaptation. The term therefore names both a specific medical-imaging algorithm and, in some summaries, a broader design pattern in which SAM provides objectness or boundary structure while task-specific modules supply class semantics, radiomics, or decoder adaptation (Alberb et al., 10 Mar 2026, Chen et al., 2023, Xiong et al., 5 Aug 2025).

1. Nomenclature and scope

The literature does not use the label uniformly. In the CRLM studies, SAMONAI is the name of the prompt propagation algorithm itself: a “zero-shot, 3D prompt propagation strategy” or “depth-aware prompt propagation algorithm” that turns the 2D, interactive Segment Anything Model into a practical volumetric segmenter for medical MRI (Alberb et al., 10 Sep 2025, Alberb et al., 10 Mar 2026). In the weakly supervised semantic segmentation summary built around SAM-enhanced pseudo-labels, “SAMONAI” appears as a thematic perspective rather than the formal title of the method; there, the operative algorithm is SEPL, which uses CAM pseudo-labels to select and combine class-agnostic SAM masks (Chen et al., 2023). In the downstream binary-segmentation baseline SAM2-UNeXT, “SAMONAI” is again used as a contextual label for a SAM-centric downstream segmentation platform rather than as a separate architecture (Xiong et al., 5 Aug 2025).

This terminological instability is reinforced by papers that explicitly redirect the query term elsewhere. One source states that “SAMONAI” corresponds to SAMO, “Sharpness-Aware Multi-task Optimization with Joint Global-Local Perturbation,” rather than to a distinct segmentation framework (Ban et al., 10 Jul 2025). Another states that the queried term likely refers to SAMOSA, “Segment Anything with Motion, GeOmetry, and Semantic Adaptation,” and that “SAMONAI” does not appear in that paper (Zhu et al., 21 May 2026). A plausible implication is that “SAMONAI” functions partly as a cross-paper shorthand for “SAM on AI” systems, but only the CRLM papers assign it a precise algorithmic definition.

2. Core algorithm in volumetric MRI

In its most concrete form, SAMONAI addresses the problem of obtaining a consistent 3D segmentation of large, concave organs such as liver and spleen from a single user point prompt per object, without finetuning SAM on medical images and without requiring a prompt on every slice (Alberb et al., 10 Sep 2025). The inputs are a 3D MRI volume, a single 2D positive point per target object on any slice and view, and optionally one background negative point. The method wraps fixed, pretrained 2D SAM inference in a three-step 3D prompt engineering and propagation procedure.

The first step is point-prompt initialization in the user-selected view. A positive point is placed inside the object on one slice; SAM is run on that slice to produce a segmentation mask and logits. This initial 2D mask appears in each orthogonal view as a line of positive points, defined by the intersection of the 3D object with those slice planes. The second step is orthogonal view seeding with optimal prompt selection. For sagittal and coronal views, SAMONAI identifies the slice where the intersection line has the maximum number of positive points, uses that slice as the seed slice, and chooses the optimal positive prompt by minimizing a weighted cost that prefers points that are central, intensity-consistent, and locally homogeneous. The third step is full-volume propagation and fusion across views. For axial, sagittal, and coronal views, candidate positive points are precomputed for every slice intersecting the object, one positive prompt per slice is selected using the same cost, a negative prompt is chosen from background points outside a 3D bounding box fitted over the three perpendicular positive lines, and 2D SAM is run slice-by-slice. To control runtime, slices are sampled at a rate of $1/3$ and missing slices are linearly interpolated (Alberb et al., 10 Sep 2025).

The prompt-selection functional is explicit. For a candidate point set $P$, SAMONAI uses

$$C_l(p) = \left\| p - \frac{1}{|P|}\sum_{q \in P} q \right\|,$$

$$C_i(p) = \left| I(p) - \operatorname{median}_{q \in P}(I(q)) \right|,$$

$$C_h(p) = \sqrt{\frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)}\big(I(r) - \mu_p\big)^2},\quad \mu_p = \frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)} I(r),$$

and the min–max normalized weighted total cost

$$C_t(p) = \alpha\,\hat{C}_l(p) + \beta\,\hat{C}_i(p) + \gamma\,\hat{C}_h(p),$$

with $\alpha=1,\ \beta=1,\ \gamma=2$, so that the optimal prompt is

$p^* = \arg\min_{p\in P} C_t(p).$

Negative prompts are restricted by the intensity threshold

$$T_N = \min(I) + 0.1\cdot(\max(I)-\min(I)),\quad I(\text{background}) \ge T_N.$$

This weighting makes homogeneity dominant and is intended to prefer stable interior prompts over edge-adjacent or intensity-ambiguous locations (Alberb et al., 10 Sep 2025).

View fusion is likewise explicit. If $L_{\text{ax}}, L_{\text{sg}}, L_{\text{co}}$ are the logits in axial, sagittal, and coronal views, then

$P$0

Binarization is conservative: $P$1 This thresholding favors high-confidence voxels and is explicitly motivated by minimizing false positives when excluding extra-hepatic tissue during liver or spleen segmentation (Alberb et al., 10 Sep 2025).

Computationally, the method performs $P$2 2D SAM calls, where $P$3 is the number of sampled slices per view, with memory dominated by storing view-wise logits and the fused volume. The slice-sampling strategy preserves near-single-view computational cost while injecting 3D awareness, and the per-slice operations are embarrassingly parallel over slices and views (Alberb et al., 10 Sep 2025).

3. Role in CRLM segmentation and survival prediction

SAMONAI was introduced to solve a concrete annotation bottleneck in CRLM imaging: tumors were annotated across the dataset, but only a small subset had liver and spleen masks. In one account, SAMONAI is used to generate liver and spleen labels for 20 cases to complete missing annotations; these masks are combined with manual tumor masks, a few-shot UNETR with a ViT Adapter is finetuned on the subset, and the resulting model is then run over the whole dataset to fill in the remaining liver and spleen labels (Alberb et al., 10 Sep 2025). In another account, the same overall strategy is described as pseudo-label generation on a 15 training / 5 validation subset, followed by fine-tuning a 3D UNETR and then retraining a final supervised UNETR with real CRLM masks and organ masks (Alberb et al., 10 Mar 2026). The common structure is a three-stage pipeline: prompt-based organ completion, few-shot organ model training, and final supervised multi-class segmentation.

The clinical framework operates on pre- and post-contrast T1-weighted MRI. For evaluation of pre-contrast organ segmentation, post-contrast manual labels were registered to pre-contrast using affine plus deformable registration (Alberb et al., 10 Sep 2025). Once organ masks are available, the segmentation pipeline applies anatomy-aware post-processing, notably extra-hepatic removal using the liver mask and small-object removal for predicted tumors with volume $P$4 (Alberb et al., 10 Sep 2025, Alberb et al., 10 Mar 2026). This anatomy-aware filtering is integral to the downstream radiomics stage because extra-hepatic false positives can corrupt per-tumor feature extraction.

The downstream survival model is SurvAMINN, an autoencoder-based multiple instance neural network for right-censored survival prediction. Tumor-level hazards $P$5 are pooled to a patient-level hazard by LogSumExp,

$P$6

and optimized with a Cox partial likelihood

$P$7

combined with an autoencoder reconstruction loss

$P$8

Within this framework, SAMONAI-generated organ masks are not an isolated endpoint; they are a bootstrap mechanism that enables anatomically plausible tumor segmentation and, subsequently, liver-specific radiomics and survival modeling (Alberb et al., 10 Sep 2025).

4. Evaluation, reported results, and version-to-version changes

Across the two CRLM reports, SAMONAI is consistently presented as improving zero-shot organ segmentation over MedSAM and as enabling an automated segmentation-plus-radiomics pipeline on a cohort of 227 CRLM patients (Alberb et al., 10 Sep 2025, Alberb et al., 10 Mar 2026). The reported numbers, however, are not identical across versions. This suggests iterative refinement of the segmentation pipeline rather than a single frozen benchmark.

Before the table, two result patterns are especially stable. First, SAMONAI is reported to outperform MedSAM in the missing-label organ setting. One report gives mean Dice improvements of more than $P$9 for liver and more than $$C_l(p) = \left\| p - \frac{1}{|P|}\sum_{q \in P} q \right\|,$$0 for spleen in pre-contrast MRI, and $$C_l(p) = \left\| p - \frac{1}{|P|}\sum_{q \in P} q \right\|,$$1 for liver and $$C_l(p) = \left\| p - \frac{1}{|P|}\sum_{q \in P} q \right\|,$$2 for spleen in post-contrast MRI (Alberb et al., 10 Sep 2025). A later report gives corresponding median improvements of $$C_l(p) = \left\| p - \frac{1}{|P|}\sum_{q \in P} q \right\|,$$3 liver and $$C_l(p) = \left\| p - \frac{1}{|P|}\sum_{q \in P} q \right\|,$$4 spleen pre-contrast, and $$C_l(p) = \left\| p - \frac{1}{|P|}\sum_{q \in P} q \right\|,$$5 liver and $$C_l(p) = \left\| p - \frac{1}{|P|}\sum_{q \in P} q \right\|,$$6 spleen post-contrast (Alberb et al., 10 Mar 2026). Second, the final survival prediction performance is essentially unchanged between reports, at a C-index of $$C_l(p) = \left\| p - \frac{1}{|P|}\sum_{q \in P} q \right\|,$$7 or $$C_l(p) = \left\| p - \frac{1}{|P|}\sum_{q \in P} q \right\|,$$8 (Alberb et al., 10 Sep 2025, Alberb et al., 10 Mar 2026).

Source	Organ segmentation results	Tumor / survival results
(Alberb et al., 10 Sep 2025)	Pre-contrast liver $$C_l(p) = \left\| p - \frac{1}{|P|}\sum_{q \in P} q \right\|,$$9, spleen $$C_i(p) = \left| I(p) - \operatorname{median}_{q \in P}(I(q)) \right|,$$0; post-contrast liver $$C_i(p) = \left| I(p) - \operatorname{median}_{q \in P}(I(q)) \right|,$$1, spleen $$C_i(p) = \left| I(p) - \operatorname{median}_{q \in P}(I(q)) \right|,$$2	Tumor Dice $$C_i(p) = \left| I(p) - \operatorname{median}_{q \in P}(I(q)) \right|,$$3 pre, $$C_i(p) = \left| I(p) - \operatorname{median}_{q \in P}(I(q)) \right|,$$4 post; F1 $$C_i(p) = \left| I(p) - \operatorname{median}_{q \in P}(I(q)) \right|,$$5 pre, $$C_i(p) = \left| I(p) - \operatorname{median}_{q \in P}(I(q)) \right|,$$6 post; C-index $$C_i(p) = \left| I(p) - \operatorname{median}_{q \in P}(I(q)) \right|,$$7
(Alberb et al., 10 Mar 2026)	Pre-contrast liver $$C_i(p) = \left| I(p) - \operatorname{median}_{q \in P}(I(q)) \right|,$$8, spleen $$C_i(p) = \left| I(p) - \operatorname{median}_{q \in P}(I(q)) \right|,$$9; post-contrast liver $$C_h(p) = \sqrt{\frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)}\big(I(r) - \mu_p\big)^2},\quad \mu_p = \frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)} I(r),$$0, spleen $$C_h(p) = \sqrt{\frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)}\big(I(r) - \mu_p\big)^2},\quad \mu_p = \frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)} I(r),$$1	Tumor Dice $$C_h(p) = \sqrt{\frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)}\big(I(r) - \mu_p\big)^2},\quad \mu_p = \frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)} I(r),$$2 pre, $$C_h(p) = \sqrt{\frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)}\big(I(r) - \mu_p\big)^2},\quad \mu_p = \frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)} I(r),$$3 post; F1 $$C_h(p) = \sqrt{\frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)}\big(I(r) - \mu_p\big)^2},\quad \mu_p = \frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)} I(r),$$4 pre, $$C_h(p) = \sqrt{\frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)}\big(I(r) - \mu_p\big)^2},\quad \mu_p = \frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)} I(r),$$5 post; C-index $$C_h(p) = \sqrt{\frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)}\big(I(r) - \mu_p\big)^2},\quad \mu_p = \frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)} I(r),$$6

The CRLM papers also report detection statistics using one-to-one matching with Dice $$C_h(p) = \sqrt{\frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)}\big(I(r) - \mu_p\big)^2},\quad \mu_p = \frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)} I(r),$$7. In the earlier account, pre-contrast performance is $$C_h(p) = \sqrt{\frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)}\big(I(r) - \mu_p\big)^2},\quad \mu_p = \frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)} I(r),$$8, and post-contrast performance is $$C_h(p) = \sqrt{\frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)}\big(I(r) - \mu_p\big)^2},\quad \mu_p = \frac{1}{|N|}\sum_{r \in N_{11\times 11}(p)} I(r),$$9, corresponding to F1-scores of $$C_t(p) = \alpha\,\hat{C}_l(p) + \beta\,\hat{C}_i(p) + \gamma\,\hat{C}_h(p),$$0 and $$C_t(p) = \alpha\,\hat{C}_l(p) + \beta\,\hat{C}_i(p) + \gamma\,\hat{C}_h(p),$$1, respectively (Alberb et al., 10 Sep 2025). That report further notes that expert review indicated many false positives and false negatives were artifactual or benign, suggesting actual precision and recall around $$C_t(p) = \alpha\,\hat{C}_l(p) + \beta\,\hat{C}_i(p) + \gamma\,\hat{C}_h(p),$$2 after correction (Alberb et al., 10 Sep 2025). The later report retains the same F1 values while reporting a higher tumor Dice, again indicating that the segmentation stack evolved while the overall clinical operating point remained similar (Alberb et al., 10 Mar 2026).

5. Broader SAM-centric usage in weak supervision and downstream segmentation

Outside medical MRI, “SAMONAI” is also used as a broader theme for exploiting SAM as an objectness prior. In weakly supervised semantic segmentation, the central idea is that CAM-derived pseudo-labels are class-aware but not object-aware: they highlight discriminative regions but often exhibit partial activation and false activation. The SEPL procedure addresses this by using CAM pseudo-labels as cues to select and combine SAM masks, thereby producing pseudo-labels that are both class-aware and object-aware (Chen et al., 2023). The pipeline is explicit: train a classifier with image-level labels $$C_t(p) = \alpha\,\hat{C}_l(p) + \beta\,\hat{C}_i(p) + \gamma\,\hat{C}_h(p),$$3, compute CAMs $$C_t(p) = \alpha\,\hat{C}_l(p) + \beta\,\hat{C}_i(p) + \gamma\,\hat{C}_h(p),$$4, produce pseudo-labels $$C_t(p) = \alpha\,\hat{C}_l(p) + \beta\,\hat{C}_i(p) + \gamma\,\hat{C}_h(p),$$5, run SAM with automatic mask proposal using $$C_t(p) = \alpha\,\hat{C}_l(p) + \beta\,\hat{C}_i(p) + \gamma\,\hat{C}_h(p),$$6, $$C_t(p) = \alpha\,\hat{C}_l(p) + \beta\,\hat{C}_i(p) + \gamma\,\hat{C}_h(p),$$7, and ViT-H, assign each SAM mask $$C_t(p) = \alpha\,\hat{C}_l(p) + \beta\,\hat{C}_i(p) + \gamma\,\hat{C}_h(p),$$8 to the class with largest intersection, and retain a mask if

$$C_t(p) = \alpha\,\hat{C}_l(p) + \beta\,\hat{C}_i(p) + \gamma\,\hat{C}_h(p),$$9

with $\alpha=1,\ \beta=1,\ \gamma=2$0 and $\alpha=1,\ \beta=1,\ \gamma=2$1. Selected masks are merged by OR to form $\alpha=1,\ \beta=1,\ \gamma=2$2. Averaged over 11 WSSS methods, this yields pseudo-label mIoU gains of $\alpha=1,\ \beta=1,\ \gamma=2$3 on VOC train and $\alpha=1,\ \beta=1,\ \gamma=2$4 on COCO train, with downstream DeepLab V2 improvements across multiple baselines (Chen et al., 2023).

A related but distinct SAM-centric usage appears in SAM2-UNeXT, where “SAMONAI” denotes a downstream segmentation platform built around encoder strengthening rather than prompt propagation. The method combines a frozen SAM2 Hiera-L encoder with adapters, a frozen DINOv2-L auxiliary encoder, dual-resolution inputs of $\alpha=1,\ \beta=1,\ \gamma=2$5 for SAM2 and $\alpha=1,\ \beta=1,\ \gamma=2$6 for DINOv2, a dense glue layer that aligns and fuses features, and a simple U-Net-style decoder (Xiong et al., 5 Aug 2025). The reported tasks are binary segmentation benchmarks including dichotomous image segmentation, camouflaged object detection, marine animal segmentation, and remote sensing saliency detection. Representative results include MAS3K $\alpha=1,\ \beta=1,\ \gamma=2$7, RMAS $\alpha=1,\ \beta=1,\ \gamma=2$8, CHAMELEON $\alpha=1,\ \beta=1,\ \gamma=2$9, and EORSSD $p^* = \arg\min_{p\in P} C_t(p).$0 (Xiong et al., 5 Aug 2025). In this usage, SAMONAI is not a named algorithm but a system-level recipe in which SAM-derived features, auxiliary semantics, and lightweight decoding jointly define the downstream segmenter.

These broader usages share a common structural motif: SAM supplies a class-agnostic structural prior, while the task-specific pipeline supplies semantics, supervision, or fusion logic. This motif is stated directly in the weak-supervision account as a broader “Segment Anything On AI” direction in which “SAM provides the objectness prior and task-specific models supply class semantics” (Chen et al., 2023).

6. Limitations, failure modes, and common misconceptions

In the medical formulation, SAMONAI is training-free and modality-agnostic, but it remains sensitive to the correctness of the initial point prompt and the quality of the seed segmentations. Misplacement of the user point, ambiguous object boundaries, post-hepatectomy margins, biliary dilation, and complex topology can misguide propagation (Alberb et al., 10 Sep 2025). The algorithm has no explicit uncertainty map or iterative refinement; confidence is implicit in conservative thresholding. It also inherits the limitations of using a 2D foundation model in a 3D setting: three-view fusion injects depth cues, but there is no full 3D reasoning, so complex motion, artifacts, or very small low-contrast structures remain challenging (Alberb et al., 10 Sep 2025). The pipeline therefore delegates tumors to a trained 3D model and applies downstream small-object removal.

In the weakly supervised formulation, failure modes arise from incorrect or missing CAM cues, SAM masks that miss parts or merge multiple objects, and hierarchical mask overlaps in which large masks engulf smaller ones (Chen et al., 2023). Because SAM mask selection is anchored on CAM overlap, object-aware refinement can worsen labels if the class cues are already wrong. These are not merely implementation details; they define the regime in which SAM-derived structural priors can be trusted.

A common misconception is that SAMONAI denotes a single standardized architecture. The literature instead contains at least three distinct usages: a specific 3D prompt propagation algorithm for medical MRI (Alberb et al., 10 Sep 2025), a broader SAM-centric pseudo-labeling perspective for weak supervision (Chen et al., 2023), and a downstream platform label for encoder-centric segmentation adaptation (Xiong et al., 5 Aug 2025). Other papers explicitly state that the queried term may instead be a misnaming of SAMO or SAMOSA rather than a separate method (Ban et al., 10 Jul 2025, Zhu et al., 21 May 2026). The most defensible encyclopedia definition is therefore narrow: SAMONAI is, first and foremost, the CRLM-oriented 3D prompt propagation algorithm; broader uses describe a family resemblance among SAM-based systems rather than a single canonical model.